RBDF: Reciprocal Bidirectional Framework for Visible Infrared Person Reidentification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Visible infrared person reidentification (VI-REID) is crucial for night-time surveillance.
• Existing methods often focus on shared features, overlooking distinct image discrepancies between modalities.

Purpose of the Study:

• To propose a novel reciprocal bidirectional framework (RBDF) for effective VI-REID.
• To address the challenge of heterogeneous image-level discrepancies in cross-modality person reidentification.

Main Methods:

• Developed bidirectional image translation subnetworks for opposite modality mappings.
• Introduced a novel associated loss function to align latent space distributions.
• Employed an attention mechanism-based feature embedding network for discriminative learning.

Main Results:

• Achieved state-of-the-art performance on the SYSU-MM01 dataset.
• Obtained 54.41% mAP and 57.66% rank-1 accuracy.
• Demonstrated significant improvements over existing VI-REID methods.

Conclusions:

• The proposed RBDF effectively unifies modalities before feature learning.
• The framework successfully eliminates the modality gap for enhanced reidentification.
• RBDF offers a promising solution for robust night-time surveillance applications.